feat: add multimodal image marker support with Ollama vision

src/gateway/mod.rs

@@ -10,7 +10,7 @@
 use crate::channels::{Channel, LinqChannel, SendMessage, WhatsAppChannel};
 use crate::config::Config;
 use crate::memory::{self, Memory, MemoryCategory};
-use crate::providers::{self, Provider};
+use crate::providers::{self, ChatMessage, Provider, ProviderCapabilityError};
 use crate::runtime;
 use crate::security::pairing::{constant_time_eq, is_public_bind, PairingGuard};
 use crate::security::SecurityPolicy;
@@ -666,6 +666,52 @@ async fn persist_pairing_tokens(config: Arc<Mutex<Config>>, pairing: &PairingGua
     Ok(())
 }
 
+async fn run_gateway_chat_with_multimodal(
+    state: &AppState,
+    provider_label: &str,
+    message: &str,
+) -> anyhow::Result<String> {
+    let user_messages = vec![ChatMessage::user(message)];
+    let image_marker_count = crate::multimodal::count_image_markers(&user_messages);
+    if image_marker_count > 0 && !state.provider.supports_vision() {
+        return Err(ProviderCapabilityError {
+            provider: provider_label.to_string(),
+            capability: "vision".to_string(),
+            message: format!(
+                "received {image_marker_count} image marker(s), but this provider does not support vision input"
+            ),
+        }
+        .into());
+    }
+
+    // Keep webhook/gateway prompts aligned with channel behavior by injecting
+    // workspace-aware system context before model invocation.
+    let system_prompt = {
+        let config_guard = state.config.lock();
+        crate::channels::build_system_prompt(
+            &config_guard.workspace_dir,
+            &state.model,
+            &[], // tools - empty for simple chat
+            &[], // skills
+            Some(&config_guard.identity),
+            None, // bootstrap_max_chars - use default
+        )
+    };
+
+    let mut messages = Vec::with_capacity(1 + user_messages.len());
+    messages.push(ChatMessage::system(system_prompt));
+    messages.extend(user_messages);
+
+    let multimodal_config = state.config.lock().multimodal.clone();
+    let prepared =
+        crate::multimodal::prepare_messages_for_provider(&messages, &multimodal_config).await?;
+
+    state
+        .provider
+        .chat_with_history(&prepared.messages, &state.model, state.temperature)
+        .await
+}
+
 /// Webhook request body
 #[derive(serde::Deserialize)]
 pub struct WebhookBody {
@@ -787,30 +833,7 @@ async fn handle_webhook(
             messages_count: 1,
         });
 
-    // Build system prompt with workspace context (IDENTITY.md, AGENTS.md, etc.)
-    let system_prompt = {
-        let config_guard = state.config.lock();
-        crate::channels::build_system_prompt(
-            &config_guard.workspace_dir,
-            &state.model,
-            &[], // tools - empty for simple chat
-            &[], // skills
-            Some(&config_guard.identity),
-            None, // bootstrap_max_chars - use default
-        )
-    };
-
-    // Call the LLM with separate system prompt
-    match state
-        .provider
-        .chat_with_system(
-            Some(&system_prompt),
-            message,
-            &state.model,
-            state.temperature,
-        )
-        .await
-    {
+    match run_gateway_chat_with_multimodal(&state, &provider_label, message).await {
         Ok(response) => {
             let duration = started_at.elapsed();
             state
@@ -994,6 +1017,12 @@ async fn handle_whatsapp_message(
     }
 
     // Process each message
+    let provider_label = state
+        .config
+        .lock()
+        .default_provider
+        .clone()
+        .unwrap_or_else(|| "unknown".to_string());
     for msg in &messages {
         tracing::info!(
             "WhatsApp message from {}: {}",
@@ -1010,30 +1039,7 @@ async fn handle_whatsapp_message(
                 .await;
         }
 
-        // Build system prompt with workspace context (IDENTITY.md, AGENTS.md, etc.)
-        let system_prompt = {
-            let config_guard = state.config.lock();
-            crate::channels::build_system_prompt(
-                &config_guard.workspace_dir,
-                &state.model,
-                &[], // tools - empty for simple chat
-                &[], // skills
-                Some(&config_guard.identity),
-                None, // bootstrap_max_chars - use default
-            )
-        };
-
-        // Call the LLM with separate system prompt
-        match state
-            .provider
-            .chat_with_system(
-                Some(&system_prompt),
-                &msg.content,
-                &state.model,
-                state.temperature,
-            )
-            .await
-        {
+        match run_gateway_chat_with_multimodal(&state, &provider_label, &msg.content).await {
             Ok(response) => {
                 // Send reply via WhatsApp
                 if let Err(e) = wa
@@ -1124,6 +1130,12 @@ async fn handle_linq_webhook(
     }
 
     // Process each message
+    let provider_label = state
+        .config
+        .lock()
+        .default_provider
+        .clone()
+        .unwrap_or_else(|| "unknown".to_string());
     for msg in &messages {
         tracing::info!(
             "Linq message from {}: {}",
@@ -1141,11 +1153,7 @@ async fn handle_linq_webhook(
         }
 
         // Call the LLM
-        match state
-            .provider
-            .simple_chat(&msg.content, &state.model, state.temperature)
-            .await
-        {
+        match run_gateway_chat_with_multimodal(&state, &provider_label, &msg.content).await {
             Ok(response) => {
                 // Send reply via Linq
                 if let Err(e) = linq